GenUI for OpenAI Build Week

A few weeks ago, OpenAI ran Build Week, a week-long hackathon hosted on Devpost to celebrate the launch of the GPT-5.6 model family. I was excited for the chance to kick the tires on GPT-5.6 and try out the recent updates to the Codex harness. I wanted a project that would let me hand real work to subagents and see how much a week of that could actually produce.

I also had unfinished business. At an in-person OpenAI hackathon back in February, I built out an idea I'd been carrying around for a while: a native SwiftUI version of Vercel's json-render. I was never able to get it where I wanted it with the models and tools at the time. Build Week seemed like the perfect excuse to resurrect it and see how far the idea could go with these new releases.

The result is GenUI, a Swift package that lets a model generate native SwiftUI interfaces.

Most AI chat apps stop at text or Markdown. I wanted to see what it would look like if an agent could answer with real native UI instead: a dashboard, a form, a comparison table, or controls that feed back into the conversation. The model never writes or runs Swift code. It produces a JSON document in a bounded format that the app validates locally and renders with a catalog of 30 native components. The host app stays in control of actions, resources, and anything sensitive.

The first half of the week went to foundations: the protocol and renderer, validation, adaptive layouts, and a live JSON workbench for iterating on documents. From there I added the runtime pieces (state management and forms) and a searchable gallery of examples covering the whole component catalog.

The second half went to the part I couldn't have done in February: teaching models to speak the format well. I built quality gates and semantic validation to reject weak outputs, then generated a corpus of 5,000 verified examples (4,000 UI generations and 1,000 repairs) and published it on Hugging Face under CC BY 4.0. I spent the final days building a complete agent chat flow on top of the renderer, with streaming and deterministic error recovery.

What made the week feel different was how the work got done. A parent Codex task coordinated 26 specialist runs, with GPT-5.6 Sol implementing features, writing tests, running the app in the Simulator, and reviewing dataset records. My job shifted from writing every line to setting direction, reviewing the work, and deciding what was good enough to keep. That division of labor held up better than I expected across 119 commits.

If you want to dig in, you can watch the 84-second demo, read the full Build Week devlog, or explore the public compiler corpus.

I'm going to keep working on it. Next up: protocol versioning, making GenUI easier to drop into another app, and fine-tuning smaller specialist models against the evaluation harness.